Keep the Coordination, Cut the Theater

Refusing to copy the company doesn't mean throwing out organizational experience wholesale. The mechanisms in human organizations fall into two classes: those that handle division of labor and those that handle theater. Keep the first. Cut the second.

Theater is not just meetings and reports. It includes process used to manufacture a sense of safety, signatures used to dilute responsibility, hierarchy used to override facts, attitude standing in for output. In human organizations these mechanisms aren't entirely useless — they often keep an organization limping along when information is incomplete, trust is scarce, and power is uneven. But an agent team should not inherit those old conditions.

Keep the Mechanisms That Handle Division of Labor

As long as multiple roles work together, the mechanisms that handle division of labor don't go away. Swap humans for agents and the material changes; the constraints remain.

Start with role boundaries. Unclear boundaries don't stop being a problem just because agents have no feelings. A planning agent that also owns execution and review will go easy on its own plan. An execution agent that doesn't know which decisions must be escalated will oscillate between timidity and overreach. A review agent with no explicit veto criteria degenerates into polite commentary. Chapter 3 covers the JD for a single agent; at the team level, the relationships between JDs must be written down too.

Then handoff contracts. Many failures happen at the seam between two agents. Upstream output becomes downstream input, and if format, scope, confidence, and open questions aren't spelled out, the downstream agent generates onward from a vague input and amplifies the error all the way down. A deliverable should be designed like an interface: what is the conclusion, what is the evidence, where is the uncertainty, which actions are forbidden, and under what conditions must it be sent back. Companies call this process; engineering calls it a contract. The name doesn't matter. The constraint does.

Then delegation and escalation paths. A group of agents can't ask a human about everything — that just hangs many nodes off one human bottleneck. Nor can every agent decide irreversible actions on its own — that's loss of control. Which decisions the role handles to completion, which get passed up, and what triggers the escalation: these rules handle division of labor, not human nature. They stay.

These three are the skeleton of collaboration, not decoration. They also serve truth-seeking: boundaries keep responsibility clear, handoffs keep the evidence chain unbroken, delegation puts a threshold in front of risk. A team without even these has no claim to facts — only mutual generation.

Cut the Organizational Theater

Human organizations carry another class of mechanism, built to cope with human self-interest, face, fear, and appetite for power: guarding against credit-grabbing, blame-shifting, slacking, factionalism, and the boss who fakes understanding and forces the call anyway. That premise does not transfer to agents.

Don't let theater count as performance. Many human performance systems evaluate people under incomplete information: read the reports, judge the attitude, notice who explains well in meetings, reward whoever packages their contribution best. An agent team doesn't need this. Every tool call, intermediate artifact, rework path, and final result can be recorded. The work record is already queryable — don't invent a ritual where the agent narrates its diligence. Performance comes from traces, artifacts, and evals, not from reporting rhetoric. The more complex the system, the more you should read the real work record and the less you should read self-description.

Don't treat approval as responsibility. Layered approval is part coordination and part responsibility anxiety — with more signatures, a failure looks like it belongs to more than one person. Having several agents nod does not mean responsibility has been handled. Several agents confirming each other does not make the facts more reliable. When approval merely spreads responsibility without changing the risk, it manufactures the comfort of "everyone looked at it" — cut it. What works is a delegation boundary and a final point of accountability: which actions run automatically, which must be escalated, and who holds the final veto.

Don't let hierarchy override facts. This is one of the hardest problems in human organizations: the people who understand have no authority to decide, the people with authority don't understand, and facts yield when they meet rank. AI does not dissolve this automatically. Design an agent team as a rigid hierarchy where the superior agent always overrides the subordinate, and you've just reskinned the old problem. Judgments of fact should flow to the role closest to the facts — the agent that just ran the trace, saw the error, and holds the context may understand that corner better than its nominal superior. Whether to accept its judgment is a question of responsibility; but rank alone cannot deny its advantage on the local facts.

Don't treat harmony as collaboration. Human teams often mistake "everyone agrees" for progress, and agent teams do it even more easily. Several agents saying "looks fine" does not show they judged independently. They may simply share the same context and be writing along the same hint. Collaboration is not mutual nodding. Collaboration is each role speaking from its own evidence and its own responsibility.

Cutting these patches has nothing to do with agents being noble. The causes have changed; the old problems must not be carried in unaltered, and the new failure modes need new mechanisms.